Claude Mods: Flipping the Harness Inside Out

What it claims

Claude Mods are Claude Code plugins made of TypeScript or JavaScript functions. They run inside the CLI and Desktop app and respond to events such as a submitted prompt, a tool call, an interface render or the end of a turn. The author argues that mods turn the harness into an "Attention Interface": a layer the user programs to decide when the agent interrupts, what it shows and which decisions it asks for. The author also forecasts that harnesses will become platforms other companies build on, much as the browser did.

How it works

A mod receives an event and can do one of three things: observe it, change how it is rendered, or handle it in place of the default behavior.

flowchart LR
  E[Event: prompt, tool call, render, turn end] --> M[Mod function]
  M -->|observe| D[Default behavior runs]
  M -->|modify| R[Changed rendering]
  M -->|handle| H[Replaces default]

Claude Code includes a built-in skill (plugin-authoring) that lets Claude write mods from a plain-language request. Running /plugin and opening Installed lists six built-in mods under Built-in:

• agents-md: loads AGENTS.md as project instructions.
• diff: takes over /diff and draws its pane.
• plugin-authoring: the mod-writing skill.
• sec-default: protects organization-managed settings from user-installed mods.
• telemetry: sends analytics records logged by Claude Code and its built-in mods.
• you-should-know: a side agent that posts a note above the prompt when it finds something easy to miss. It is the only built-in that is off by default.

In the author's example, You Should Know warned that CompareView.swift and about 45 other files held uncommitted work before a demo edit. It suggested a commit or stash so the earlier work would not be discarded.

The author's Next Steps Supervisor ships in ModSmith, an open-source plugin with several mods and a template for building new ones that work together. It runs after any turn that edited files, ran commands or made 8 or more tool calls. It forks the conversation, reviews it, lists flagged items and offers a suggested next step as a clickable button that sends the instruction to Claude Code. The page names ModSmith but gives no install command.

Evidence

Since installing the supervisor on October 15, 99 of the author's 111 Claude Code turns met its trigger. The ModSmith README estimates each check costs about 5,000 to 6,000 input-equivalent tokens on a 50,000-token session when served from the prompt cache. It costs more when the cache has expired.

Limits

• Noise: the supervisor fires on nearly every turn, and the author reports skimming past it. In one session it marked a task "Solved" while also listing two gaps and three shortcuts. /supervisor off disables it for a session.
• Trust: Anthropic's docs state that "mods aren't sandboxed." A loaded mod can read and write files, see every prompt and spend usage. Built-in mods keep running under --safe-mode. The author reads third-party mod code before installing it.

For comparison, Cursor's PR Routing & Approval reads approval policy from the base branch when a PR edits that policy. This means an agent cannot relax review on its own PR. Cursor's docs say the feature "does not replace a full code review." The page describes no equivalent guard for mods. It reports ChatGPT Intelligent UI as shipping with GPT-6 on October 7, with a control to turn visuals down as the one user control mentioned at launch.